13-Benchmark SOTA! Mininglamp Technology Officially Open-Sources GUI-VLA Model Mano-P 1.0

BEIJING, April 15, 2026 /PRNewswire/ -- Mininglamp Technology has officially open-sourced Mano-P 1.0, a self-developed GUI-aware agent model capable of executing complex cross-platform tasks entirely through pure vision. By seamlessly controlling desktop software, web interfaces, and complex graphical workflows, Mano-P breaks the "see but not do" limitation of AI, empowering users to build personalized AI while guaranteeing data sovereignty and ushering in the era of privatized personal AI.

Breaking the "See but Not Do" Barrier with Pure Vision

Unlike traditional automation constrained by API calls or HTML parsing, Mano-P utilizes pure visual understanding to operate without external interfaces. It serves as a critical execution foundation for the Agent ecosystem, integrating seamlessly into agents like OpenClaw. This breakthrough eliminates manual intervention bottlenecks, enabling autonomous execution in complex commercial scenarios such as 3D applications and professional tools that were previously inaccessible to standard agents.

Unrivaled Performance Across Global Benchmarks

The model redefines performance ceilings, achieving overwhelming State-of-the-Art (SOTA) results across 13 authoritative multimodal benchmarks. Operating on a dual-version architecture, the 72B model ranked first globally in the OSWorld benchmark with a 58.2% task success rate, leading the runner-up by 13.2 percentage points, while also dominating systems like ScreenSpot-V2 and MMBench. Furthermore, the highly efficient 4B quantized model supports local execution on Apple M4 Pro devices, achieving up to 476 tokens/s prefill speeds with just a 4.3GB memory footprint.

Local Deployment for Enterprise-Grade Privacy

As AI integrates into core workflows, data privacy is paramount. Mano-P ensures zero data uploads to the cloud by running locally on Mac devices or via a dedicated USB 4.0 compute stick. This "pure visual understanding plus local execution" architecture guarantees physical isolation from external networks, allowing AI to securely drive complex business processes and self-correct even in offline, high-privacy enterprise environments where data sovereignty is a non-negotiable requirement.

Fostering a Prosperous Personalized AI Ecosystem

Released under the Apache 2.0 license, Mano-P provides out-of-the-box usage modes, supporting commercial use and secondary development without complex API configurations. Mininglamp Technology is initially open-sourcing the Mano-CUA core skill for easy integration into existing workflows like OpenClaw or Claude Code, with local models and SDKs following within the month. Future releases will include underlying training methods and token pruning technology, accelerating the global development of a personalized AI ecosystem.

SOURCE Mininglamp Technology